From 56a357ed33dec39357344816ccba2732109a96a8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Nicol=C3=B2=20Lucchesi?= Date: Mon, 13 Jul 2026 10:16:24 +0200 Subject: [PATCH] [Bugfix][KV Cache] Don't route uniform-page-size MLA+SWA models into DeepseekV4 packing (#48256) Signed-off-by: NickLucche Co-authored-by: Claude --- tests/v1/core/test_kv_cache_utils.py | 43 ++++++++++++++++++++++++++++ vllm/v1/core/kv_cache_utils.py | 5 ++++ 2 files changed, 48 insertions(+) diff --git a/tests/v1/core/test_kv_cache_utils.py b/tests/v1/core/test_kv_cache_utils.py index 0c212cb089..e7520b1ee9 100644 --- a/tests/v1/core/test_kv_cache_utils.py +++ b/tests/v1/core/test_kv_cache_utils.py @@ -32,6 +32,7 @@ from vllm.v1.core.kv_cache_utils import ( get_kv_cache_configs, get_max_concurrency_for_kv_cache_config, get_request_block_hasher, + group_and_unify_kv_cache_specs, hash_block_tokens, init_none_hash, is_kv_cache_spec_uniform, @@ -1951,6 +1952,48 @@ def new_mla_spec(cache_dtype_str=None): ) +def new_swa_mla_spec(head_size=576, sliding_window=128): + return SlidingWindowMLASpec( + block_size=16, + num_kv_heads=1, + head_size=head_size, + dtype=torch.float32, + sliding_window=sliding_window, + ) + + +def test_group_and_unify_kv_cache_specs_no_swa_mla_returns_none(): + # Without any SlidingWindowMLASpec the function does not apply. + specs = {"mla.0": new_mla_spec(), "mla.1": new_mla_spec()} + assert group_and_unify_kv_cache_specs(specs) is None + + +def test_group_and_unify_kv_cache_specs_uniform_page_size_returns_none(): + # A non-DeepseekV4 model that mixes full MLA and sliding-window MLA layers + # with a uniform page size must not fall into the DeepseekV4 tuple-packing + # path; it should defer to the generic uniform-page-size grouping instead. + mla_spec = new_mla_spec() + swa_spec = new_swa_mla_spec() + assert mla_spec.page_size_bytes == swa_spec.page_size_bytes + specs = {"mla.0": mla_spec, "mla.1": new_mla_spec(), "swa.0": swa_spec} + assert group_and_unify_kv_cache_specs(specs) is None + + +def test_group_and_unify_kv_cache_specs_mixed_page_size_groups(): + # DeepseekV4-style: differing page sizes across MLA and sliding-window MLA + # layers do require tuple packing, so grouping must still be produced. + mla_spec = new_mla_spec() + swa_spec = new_swa_mla_spec(head_size=1024) + assert mla_spec.page_size_bytes != swa_spec.page_size_bytes + specs = {"mla.0": mla_spec, "mla.1": new_mla_spec(), "swa.0": swa_spec} + grouped = group_and_unify_kv_cache_specs(specs) + assert grouped is not None + # One MLA group plus one sliding-window MLA group. + assert len(grouped) == 2 + layer_names = {name for g in grouped for name in g.kv_cache_specs} + assert layer_names == {"mla.0", "mla.1", "swa.0"} + + def test_get_kv_cache_spec_kind_prefers_specific_attention_subclasses(): assert get_kv_cache_spec_kind(new_mla_spec()) == KVCacheSpecKind.MLA_ATTENTION diff --git a/vllm/v1/core/kv_cache_utils.py b/vllm/v1/core/kv_cache_utils.py index 8adc472672..dc5ecf3f4d 100644 --- a/vllm/v1/core/kv_cache_utils.py +++ b/vllm/v1/core/kv_cache_utils.py @@ -1535,6 +1535,11 @@ def group_and_unify_kv_cache_specs( ): return None + # SlidingWindowMLASpec models with uniform page sizes don't need tuple packing. + page_sizes = {spec.page_size_bytes for spec in kv_cache_spec.values()} + if len(page_sizes) <= 1: + return None + mla_specs: dict[str, KVCacheSpec] = {} grouped_swa_mla_specs: dict[tuple[int, int], dict[str, KVCacheSpec]] = defaultdict( dict